What happens when I turn off a subagent?

The Claude Code session stops at turns that role owns. Turn off Code and it halts at Write train_*.py. Optional Deploy can stay off and you still finish a prototype if required turns pass.

Why is the Orchestrator always on?

The orchestrator writes .cursor/plan.md and spawns specialists via Task(subagent=...). Claude Code sessions need one coordinator to delegate work.

How is this different from the RL agent playground?

The RL playground trains a Q-learning policy. This lab simulates Claude Code subagents, Task() delegation, and Read/Write/Bash tool use—not reinforcement learning.

Is it free with no install?

Yes. A browser-only educational simulator—no LLM API calls. Explore Claude Code session flow and subagent dependencies interactively.

Everyone's AI

NN Classifier Loss Landscape Lab KNN Neighbor Classifier RL Agent Conv Vision Attention Playground Claude Agents

Everyone's AI

NN Classifier Loss Landscape Lab KNN Neighbor Classifier RL Agent Conv Vision Attention Playground Claude Agents

Playground

Claude Code multi-agent

Follow Claude Code as it receives your prompt, spawns subagents, and solves ML tasks with Read·Write·Bash·Glob tools!

Turn0/8

SpeedShow formulas

Control help

Toggle specialists in the subagent roster. ▶ Play lets Claude Code delegate via Task(); disabled roles halt the session.

User prompt

Which ML task should Claude Code handle?

Each scenario is an ML task Claude Code would tackle in a real repo. The orchestrator delegates via Task(subagent=...), and each agent invokes tools.

Prompt

Customer churn prediction

“We have telco customer records in telco_churn.csv. Build an XGBoost classifier and reach at least AUC 0.80 on validation.”

🎯 Deliver a validated churn model with AUC ≥ 0.80

All subagents are assumed to run inside a Claude Code session.

💡 Turn Code off, press ▶ Play — the session stops at Write train_*.py.

Task decomposition workbook

8 steps · subtasks per agent

Roadmap of how the orchestrator splits the user prompt into 8 subtasks and delegates via Task(). ▶ Play highlights the current turn.

Problem to solve

Customer churn prediction

Deliver a validated churn model with AUC ≥ 0.80

1
Orchestratororchestrator·Session plan
Break the big problem into 8 subtasks
2
Datadata-agent·Data retrieval
Find the CSV dataset matching the request
3
Datadata-agent·Preprocessing
Build a trainable feature matrix
4
Modelmodel-agent·Model design
Design a classifier suited to tabular data
5
Modelmodel-agent·HPO
Find hyperparameters that maximize validation AUC
6
Codecode-agent·Code generation
Write a reproducible training script
7
Verifyverify-agent·Verification
Check AUC target and data leakage
8
Deploydeploy-agent·Deploy
Save model artifact (deploy optional)

Subagent roster

Toggle spawnable specialists ON/OFF

Claude CodeClaude Code session

Orchestrator writes plans and delegates via Task(subagent=...) · each agent uses Read/Write/Bash/Glob tools

Active subagents 5/6

Subagents that the Claude Code orchestrator invokes via Task(). The orchestrator is always active; toggle the rest.

OrchestratororchestratorClaudeRequired
Spawn subagents via Task() · write .cursor/plan.md
orchestrator · Task() spawn · .cursor/plan.md
Always onOn
Datadata-agentClaudeRequired
Glob/Read/Grep · Write/Bash preprocessing
Explore with Glob/Read/Grep · preprocess with Write/Bash
On
Modelmodel-agentClaudeRequired
Write/Edit configs · Bash HPO
Write configs/*.yaml · HPO/tuning with Bash
On
Codecode-agentClaudeRequired
Write/Edit train/serve scripts
Write/Edit train_*.py · serve_*.py
On
Verifyverify-agentClaudeRequired
Bash evaluation & audit · Read metrics
Bash pytest/bench · Read metrics.json
On
Deploydeploy-agentClaudeOptional
Write manifests · kubectl/curl Bash
Write deploy/*.yaml · kubectl/curl Bash
Off

Required for this scenario

📊 Data ✓
🧠 Model ✓
💻 Code ✓
✅ Verify ✓
🚀 Deploy (Optional)

Session board

Click to toggle · turn progress

The horizontal track shows subagent delegation order in the Claude Code session. Click a role to toggle spawn targets; the session resets.

Claude Code session

·Claude Code

You can toggle subagents from both the roster and the session board.

Claude Code session log

Task delegation · tool use · context handoff

Each turn is subagent tool use. Inspect Glob/Read/Write/Bash calls and workspace context handoff.

Claude Code~/modudl-ml · Claude CodeOrchestrator writes plans and delegates via Task(subagent=...) · each agent uses Read/Write/Bash/Glob tools

👤

Customer churn prediction

We have telco customer records in telco_churn.csv. Build an XGBoost classifier and reach at least AUC 0.80 on validation.

🎯 Deliver a validated churn model with AUC ≥ 0.80

Session progress0%

Press ▶ Play to start the Claude Code session.

Session result

Did all required turns pass?

Active subagents

Total turns

Turns done

Required roles

Press ▶ Play or Step to run the Claude Code session.

Related lessons

Playground

Claude Code multi-agent

Follow Claude Code as it receives your prompt, spawns subagents, and solves ML tasks with Read·Write·Bash·Glob tools!

Turn0/8

SpeedShow formulas

Control help

Toggle specialists in the subagent roster. ▶ Play lets Claude Code delegate via Task(); disabled roles halt the session.

User prompt

Which ML task should Claude Code handle?

Each scenario is an ML task Claude Code would tackle in a real repo. The orchestrator delegates via Task(subagent=...), and each agent invokes tools.

Prompt

Customer churn prediction

“We have telco customer records in telco_churn.csv. Build an XGBoost classifier and reach at least AUC 0.80 on validation.”

🎯 Deliver a validated churn model with AUC ≥ 0.80

All subagents are assumed to run inside a Claude Code session.

💡 Turn Code off, press ▶ Play — the session stops at Write train_*.py.

Task decomposition workbook

8 steps · subtasks per agent

Roadmap of how the orchestrator splits the user prompt into 8 subtasks and delegates via Task(). ▶ Play highlights the current turn.

Problem to solve

Customer churn prediction

Deliver a validated churn model with AUC ≥ 0.80

1
Orchestratororchestrator·Session plan
Break the big problem into 8 subtasks
2
Datadata-agent·Data retrieval
Find the CSV dataset matching the request
3
Datadata-agent·Preprocessing
Build a trainable feature matrix
4
Modelmodel-agent·Model design
Design a classifier suited to tabular data
5
Modelmodel-agent·HPO
Find hyperparameters that maximize validation AUC
6
Codecode-agent·Code generation
Write a reproducible training script
7
Verifyverify-agent·Verification
Check AUC target and data leakage
8
Deploydeploy-agent·Deploy
Save model artifact (deploy optional)

Subagent roster

Toggle spawnable specialists ON/OFF

Claude CodeClaude Code session

Orchestrator writes plans and delegates via Task(subagent=...) · each agent uses Read/Write/Bash/Glob tools

Active subagents 5/6

Subagents that the Claude Code orchestrator invokes via Task(). The orchestrator is always active; toggle the rest.

OrchestratororchestratorClaudeRequired
Spawn subagents via Task() · write .cursor/plan.md
orchestrator · Task() spawn · .cursor/plan.md
Always onOn
Datadata-agentClaudeRequired
Glob/Read/Grep · Write/Bash preprocessing
Explore with Glob/Read/Grep · preprocess with Write/Bash
On
Modelmodel-agentClaudeRequired
Write/Edit configs · Bash HPO
Write configs/*.yaml · HPO/tuning with Bash
On
Codecode-agentClaudeRequired
Write/Edit train/serve scripts
Write/Edit train_*.py · serve_*.py
On
Verifyverify-agentClaudeRequired
Bash evaluation & audit · Read metrics
Bash pytest/bench · Read metrics.json
On
Deploydeploy-agentClaudeOptional
Write manifests · kubectl/curl Bash
Write deploy/*.yaml · kubectl/curl Bash
Off

Required for this scenario

📊 Data ✓
🧠 Model ✓
💻 Code ✓
✅ Verify ✓
🚀 Deploy (Optional)

Session board

Click to toggle · turn progress

The horizontal track shows subagent delegation order in the Claude Code session. Click a role to toggle spawn targets; the session resets.

Claude Code session

·Claude Code

You can toggle subagents from both the roster and the session board.

Claude Code session log

Task delegation · tool use · context handoff

Each turn is subagent tool use. Inspect Glob/Read/Write/Bash calls and workspace context handoff.

Claude Code~/modudl-ml · Claude CodeOrchestrator writes plans and delegates via Task(subagent=...) · each agent uses Read/Write/Bash/Glob tools

👤

Customer churn prediction

We have telco customer records in telco_churn.csv. Build an XGBoost classifier and reach at least AUC 0.80 on validation.

🎯 Deliver a validated churn model with AUC ≥ 0.80

Session progress0%

Press ▶ Play to start the Claude Code session.

Session result

Did all required turns pass?

Active subagents

Total turns

Turns done

Required roles

Press ▶ Play or Step to run the Claude Code session.

Related lessons